My BladeRunner Project

A Quadrupedal Running Robot

I’m a robotics enthusiast who worked with five other cross-functional engineers to explore advanced simulation, optimization, and control systems. In this project, I designed and simulated a quadrupedal running robot equipped with innovative running blades inspired by high-performance athletes and the biomechanics of cheetahs.

Mission Objective

My primary goal was to create a fast and efficient robotic quadruped that replicates the powerful yet agile running style of cheetahs. I aimed to leverage advanced blade designs, biomimicry insights, and a passive knee mechanism to maximize speed, stability, and energy savings.

BladeRunner Project Title

Skills & Tools

Core Expertise

I combined multiple disciplines to bring the BladeRunner project to life. Here are some of the key skills and tools I used:

  • Programming: MATLAB
  • Simulation Tools: MATLAB Simscape Multibody, Simulink
  • Optimization Techniques: Genetic Algorithms, Reinforcement Learning (DDPG)
  • Design & CAD: SolidWorks, Onshape
  • Control Systems: Dynamic System Modeling, Feedback Control, Sensor Integration

Project Overview

Introduction and Motivation

My inspiration came from the extraordinary running mechanics of cheetahs and the proven efficiency of Paralympic running blades. By blending mechanical design, simulation, and optimization, I tackled challenges such as dynamic stability, joint coordination, and energy management.

Transition: After defining the project's objectives, I moved on to developing a robust simulation strategy to validate my initial designs.

Methodology

Simulation Setup and System Design

I adopted a simulation-first approach, creating a simplified CAD model in Onshape and importing it into MATLAB Simscape Multibody. This enabled me to rapidly iterate on the robot’s geometry and test various leg configurations, reducing the risk of costly physical prototyping errors.

Transition: To further refine my approach, I explored a passive knee mechanism, advanced blade designs, and biomimicry-inspired gait cycles.

Passive Knee Mechanism

Leveraging Momentum for Efficiency

One key innovation was the introduction of a passive knee mechanism in the rear legs. From my literature review, I learned that allowing the knee to lock during ground contact can significantly reduce energy consumption. Essentially, the knee travels in sync with the running motion by harnessing the leg’s momentum rather than requiring constant motor torque.

This approach minimizes power demands during the stance phase and aligns with research showing that limiting active control when ground reaction forces are highest leads to more efficient locomotion.

Passive Knee Mechanism

Blade Design

C-Shaped vs. J-Shaped Running Blades

Through a thorough literature review, I discovered that running blades generally come in two forms: C-shaped and J-shaped. C-shaped blades often cater to endurance running with moderate energy return and stability. In contrast, J-shaped blades excel at delivering high energy return at fast speeds, making them ideal for sprinting.

Given my goal of maximizing speed, I selected a J-shaped blade design for its superior ground clearance and pronounced spring effect. This choice allowed my robot to store and release energy efficiently, playing a pivotal role in achieving a fast yet stable running gait.

Blade Design

Biomimicry

Animal Gait Cycles and Speed Variations

I conducted a literature review on animal locomotion to understand how gait cycles differ among species like dogs, horses, and cheetahs. I observed that at higher speeds, many quadrupeds transition from diagonal synchronization to a front-and-back leaping motion, which maximizes ground contact time during propulsion.

Based on these insights, I began my genetic algorithm with a diagonal gait cycle while encouraging a front-and-back leaping motion in the reward function. This hybrid approach provided the flexibility to discover faster, more powerful strides, merging stability with explosive speed.

Biomimicry in Gait Cycles

Simscape Implementation

High-Fidelity Simulation in MATLAB

After refining the design choices—passive knee, J-shaped blades, and biomimicry-inspired gaits—I moved to MATLAB Simscape Multibody for detailed dynamic analysis. I tested how the blades compressed and released energy, how the passive knee responded to ground contact, and how torque inputs influenced forward velocity and stability.

Simulink and the Physical Modeling Toolbox provided real-time insights into ground reaction forces, joint stress, and center-of-mass movement. This thorough simulation phase helped validate each subsystem before integrating advanced optimization methods.

MATLAB Simscape Implementation

Genetic Algorithm

Overview of the Genetic Algorithm (GA)

To systematically refine my quadruped’s gait, I employed a genetic algorithm (GA). The process began with an input gait—a walking trajectory from literature—and defined upper and lower angle limits for each joint. I then generated a mutated population of joint angles and step timings, allowing the GA to explore various stride profiles.

  • Input Gait: Walking trajectory from literature
  • Angle Limits: Defined upper and lower bounds for each joint
  • Mutated Population: Variations in joint angles & step timings to influence stride and trajectory
  • Iterative Process: Evaluated each generation via a reward function, converging on the best gait

Rewards

  • Longitudinal (X) distance traveled

Penalties

  • Lateral (Y) displacement
  • Aggressiveness: Sign flips in joint velocity
  • Torso pitch instability: Exceeding 1 standard deviation from mean

By focusing on maximizing forward motion while minimizing instability, the GA refined my robot’s gait, striking a balance between speed and stability that aligns with my mission objective.

Algorithm Progress Videos

Below are three key milestones from my GA runs. The first video shows the starting algorithm, the second highlights the passive knee integration, and the final clip demonstrates the converged gait.

Starting Algorithm

With Passive Knee

Final Results

Reinforcement Learning

Enhancing Performance through Reinforcement Learning

In parallel with the genetic algorithm, I applied a reinforcement learning (RL) strategy using Deep Deterministic Policy Gradient (DDPG). Rather than relying on a predefined gait, the RL agent discovered optimal joint torques through continuous interaction with the simulation environment.

I used an actor-critic framework, where the agent received continuous feedback on joint angles, body orientation, and ground reaction forces. Custom penalty functions discouraged abrupt velocity changes and prolonged ground contact, resulting in a smoother, more energy-efficient gait.

Final Results

Conclusive Analysis and Achievements

The combination of a passive knee mechanism, J-shaped blades, and biomimicry-driven gait cycles significantly enhanced my quadruped’s locomotion. Through iterative refinement with genetic algorithms and reinforcement learning, I achieved a gait that allowed the robot to cover 23 meters in 10 seconds while maintaining excellent stability and energy efficiency.

Final Results

This project underscored the power of simulation-based design and optimization in robotics. I’m excited about potential real-world applications, such as faster search-and-rescue robots or agile courier systems. Collaborating with five cross-functional engineers has deepened my appreciation for teamwork, as each subsystem was carefully integrated for optimal results.

References

References and Supplementary Materials

  1. Taboga P, Beck ON, Grabowski AM (2020). Influence of prosthetic shape on maximum speed in sprinters with bilateral transtibial amputations.https://doi.org/10.1371/journal.pone.0229035
  2. Y. M. Shabana et al. (2022). Design and simulation of prosthetic running blade using functionality graded materials.https://doi.org/10.1109/NILES56402.2022.9942368
  3. Choi J. (2021). Multi-phase joint-angle trajectory generation inspired by canine motion for quadruped robot control. Sensors, 21(19), 6366.https://doi.org/10.3390/s21196366
  4. Zhang, X., Zhao, C., Xu, Z., and Huang, S. (2022). Mechanism analysis of cheetah high-speed locomotion based on digital reconstruction. Biomimetic Intelligence and Robotics, 2(1), 100033.https://doi.org/10.1016/j.birob.2021.100033

I have also compiled a more detailed report and presentation slides that dive deeper into the design decisions, simulation data, and optimization metrics used throughout this project.